Anthropic IPO: What 12 AI terms should investors know before buying shares amid $16.3 billion revenue and an $80.7 billion loss?

Anthropic has filed its IPO prospectus, and the early pages define 39 terms. Some, like “AI”, need no explanation. Others do. These words will likely come up on earnings calls and in 10-Q and 10-K filings, so it helps to know them before you invest.

Anthropic’s revenue for the first half of 2026 was $16.3 billion, up from $1.1 billion a year earlier. Its net loss for the same period was $80.7 billion, which points to heavy AI spending in the coming years.

Bankers are reportedly aiming for a valuation of $1.5 trillion to $2 trillion, and the listing could happen before the end of November. OpenAI is expected to list in 2027 and may use similar language.

Terms about how AI does the work

Agent loop: An AI agent gets a task, takes an action (such as planning, running code or calling a tool), checks the result and decides what to do next. It repeats this until the task is done.

Multi-agent orchestration: Several AI agents work on one task. One coordinates while the others handle smaller parts at the same time. This allows longer and more complex work than a single agent can manage.

Interleaved thinking: The model reasons between tool calls within one response. It does not plan only once at the start.

Context window: The most content a model can handle in one request, measured in tokens. It covers instructions, past conversation, the user’s input and the model’s own output.

Terms about training and safety

Reinforcement learning: A model learns by taking actions and getting feedback, called rewards. It adjusts to earn more rewards over time. It is widely used after pre-training to improve helpfulness, accuracy and safety.

RLHF: Short for reinforcement learning from human feedback. Human reviewers rate the model’s answers, and those ratings serve as the reward.

Constitutional AI: Anthropic trains its models using written principles, called Claude’s constitution. It has four priorities: being broadly safe, broadly ethical, compliant with Anthropic’s guidelines and genuinely helpful. The models check and revise their own answers against these principles. Anthropic says this lets it control behaviour through auditable principles, not only fixed rules or human reviewers.

Mechanistic interpretability: Researchers study the inner workings of neural networks to see what causes a model’s behaviour, instead of judging only from inputs and outputs. The goal is to catch problems, including deceptive or unsafe behaviour, that normal testing may miss.

Compute multipliers: Improvements in methods that get more output from the same amount of computing power, in both training and use. For a company spending this much on compute, such gains matter.

Terms about customers and risks

Applied AI: A team function that builds AI into a customer’s real business operations. These teams include forward-deployed engineers, who work with companies to build custom solutions around the customer’s own systems and processes.

Prosumers: Individuals who pay for AI tools to do knowledge work.

Prompt injection: An attack in which harmful instructions are hidden in content the AI reads, such as websites, documents or tool outputs. The aim is to change the model’s behaviour, usually without the user knowing.